BlackPearl Team Wins All Three Tracks of KDD 2024 OAG‑Challenge Cup with Large‑Model Solutions

ACM SIGKDD (Knowledge Discovery and Data Mining) is the premier international conference in data mining. The KDD Cup, organized by SIGKDD since 1997, is the most influential competition in the field.

Recently, the BlackPearl team from Meituan's Dazhong Dianping Technology Department / Search and Content Intelligence group participated in the KDD 2024 OAG‑Challenge Cup, tackling three tracks: WhoIsWho‑IND, PST, and AQA. The team won the champion in all three tracks with a large margin.

In the WhoIsWho (same‑name disambiguation) track, the team introduced an iterative large‑model text clustering method enhanced by self‑feedback. This approach can directly process structured information and output clustering results end‑to‑end, achieving an 83 % gAUC score that significantly surpasses traditional machine‑learning baselines.

In the PST (paper source tracing) track, the team applied graft learning to inject the powerful semantic matching capability of BERT‑like models into a large language model, boosting sample confidence. They also built a Retrieval‑Augmented Generation (RAG) based automatic feature‑engineering pipeline to mitigate noisy, multi‑modal textual data. Their 7B single model outperformed a ChatGPT + RAG solution by 10 % on the MAP metric.

For the AQA (academic paper question answering) track, the team addressed heavy noise in the data. By combining LLM‑for‑Vector and Boosting techniques, they created an integrated recall‑and‑ranking Boosting LLM for Searching solution, which outperformed traditional text‑embedding search methods and transferred the LLM’s semantic understanding to noisy academic search scenarios.

The BlackPearl team has open‑sourced all three solutions on GitHub ( https://github.com/BlackPearl-Lab/KddCup-2024-OAG-Challenge-1st-Solutions ) for the research community, and will present their findings at KDD 2024 in Barcelona. Future work will continue to explore large‑model technologies to enhance Meituan’s products and services.